GPUs
What Is NVIDIA Blackwell?
The architecture behind the B200 and GB200, and what it changes at rack scale.
Updated
Blackwell is the NVIDIA data-center architecture that followed Hopper.
Its defining engineering choice is packaging two reticle-limited dies so they behave as a single GPU, connected by a very high bandwidth die-to-die link. It also adds support for lower precision numeric formats that increase inference throughput, and it is paired with a new generation of interconnect.
The more important shift is commercial: Blackwell is sold and deployed primarily as rack-scale systems in which dozens of GPUs share a coherent high-bandwidth domain. That moves the unit of purchase from the card to the rack, and pushes liquid cooling from optional to expected.
GPU Data Hub Daily
Stay Ahead of the AI Infrastructure Economy
The most important GPU, AI, data-center, semiconductor and cloud developments delivered directly to your inbox.
Cite this page
“What Is NVIDIA Blackwell?.” GPU Data Hub. https://gpudatahub.com/guides/what-is-nvidia-blackwell
You are welcome to reference and link to this page. Please link to the URL above.